OpenAI Adds Ultrafast Mode to GPT-5.6 Sol for Real-Time AI

For frontier AI models, intelligence has usually come with an uncomfortable trade-off: the smarter the model, the longer users may have to wait for complex answers.

OpenAI is now trying to break that link.

The company has introduced Ultrafast, a new processing tier for GPT-5.6 Sol that can run the model at up to 14 times the speed of Standard processing. More strikingly, OpenAI says it can produce as many as 750 output tokens per second, putting its flagship model into territory normally associated with much smaller, latency-focused AI models.

Ultrafast is launching first through the OpenAI API and is currently available only as a limited preview for selected customers. OpenAI says access will expand as additional capacity becomes available.

The numbers explain why this update matters

OpenAI already offered a faster way to run GPT-5.6 Sol. Its existing Fast mode can deliver up to 2.5× the speed of Standard processing and costs twice as much. Ultrafast moves considerably further, although OpenAI has not yet publicly disclosed its pricing.

That 14× figure is an “up to” measurement, rather than a guarantee that every API request will complete exactly 14 times faster. Actual end-to-end performance can also depend on reasoning time, prompt size, tool calls, network latency and the type of workload being executed.

Still, moving a frontier model toward hundreds of generated tokens per second changes more than how quickly text appears on a screen.

OpenAI wants Sol inside real-time workflows

OpenAI's reasoning behind Ultrafast becomes clearer from the early applications it highlights.

A conventional AI assistant can afford to take some time producing a report. An AI system participating in a live customer call, responding to a production outage or analyzing changing financial information does not have the same luxury.

OpenAI says it is testing Ultrafast across coding, commerce, financial research, customer support and interactive applications. Early customers include Jane Street, Podium, Basis and Rogo.

Podium, for example, has been experimenting with the technology in its voice AI stack, where latency is especially noticeable. Rogo has tested it for financial research, while OpenAI itself is applying Ultrafast to incident response and internal research workflows.

During an outage, OpenAI engineers can use it to move rapidly through logs, traces and team conversations, develop a likely explanation, test it and prepare a possible fix. OpenAI is also exploring whether experiments that previously ran overnight can instead become iterative research sessions completed during the working day.

That is a much more meaningful benchmark for speed than simply asking how rapidly a chatbot can finish a paragraph.

Cerebras is behind the speed jump

The infrastructure powering Ultrafast comes from Cerebras, extending OpenAI's partnership with the AI hardware company.

OpenAI says Cerebras infrastructure is responsible for delivering GPT-5.6 Sol at up to 750 output tokens per second. The partnership is aimed at ultra-low-latency inference, where the challenge is not necessarily making the model smarter but getting an already powerful model to return its computations much faster.

This also signals an interesting shift in the AI race. Model developers have spent years competing over benchmark scores, parameter efficiency and reasoning capabilities. Inference infrastructure is increasingly becoming another battleground.

A model that is only marginally smarter but dramatically faster can be far easier to put into live software.

There is an important catch: access and pricing

Ultrafast is not yet something every GPT-5.6 user can switch on.

OpenAI describes the release as a limited preview for a select group of customers, with broader availability planned as capacity increases. The company has also not announced public Ultrafast pricing.

For comparison, normal GPT-5.6 Sol API usage currently costs $5 per million input tokens and $30 per million output tokens. OpenAI's Fast processing tier increases processing speed while charging a premium, making it reasonable to expect Ultrafast to be positioned as a premium service, although OpenAI has not yet confirmed its exact cost structure.

There is another naming distinction worth noting. Ultrafast is a processing-speed tier. It is not the same as GPT-5.6 Sol's “ultra” reasoning setting, which uses multiple agents in parallel and can actually consume more tokens in exchange for stronger results on difficult tasks.

Speed is becoming part of the AI product itself

OpenAI's Ultrafast announcement may look like an infrastructure upgrade, but its implications are broader.

GPT-5.6 already focuses heavily on completing professional and agentic work more efficiently. OpenAI reports a 92.2% BrowseComp score for Sol's multi-agent configuration and 62.6% on OSWorld 2.0, while the company has also emphasized reductions in tokens and completion time across coding and professional workflows.

Ultrafast adds another dimension: how much intelligence can be delivered per second, rather than only per token or per dollar.

If OpenAI can eventually make 750-token-per-second frontier inference broadly available at commercially workable prices, AI agents could move deeper into applications where several seconds of delay currently make them impractical.

That could make Ultrafast less about making ChatGPT answers appear faster and more about enabling a new generation of AI systems that have to think quickly enough to participate in events while they are still happening.